Back to search

IKTPLUSS-IKT og digital innovasjon

TailorMade: Tailoring Server Processors to Application Characteristics for Sustainable Warehouse-Scale Computing

Alternative title: TailorMade: Tailoring Server Processors to Application Characteristics for Sustainable Warehouse-Scale Computing

Awarded: NOK 7.9 mill.

TailorMade aims to take a leap towards sustainable large-scale computing, also called cloud computing. We propose to specialize the hardware, especially processors, used in these warehouse-scale computers to match the characteristics of applications running on them. The computing landscape is shifting from the traditional desktop computing to mobile-cloud computing model. In this new model, cloud computing has emerged as the workhorse that does all the heavy-duty computations and serves as the backbone of mobile services. However, the warehouse-scale computing is rapidly becoming unsustainable as current approaches require increasingly higher server count, and other support infrastructure, to meet the swiftly growing demands for online web-services. As a result, a typical warehouse-scale computer, also called datacenter, already houses tens-of-thousands of servers, consumes tens of Megawatts of power, and costs hundreds-of-millions of dollars. Ignoring other aspects, even the power consumption itself is increasing at an unsustainable rate with datacenters projected to consume about 8% of world?s energy by 2030, compared to 1% in 2010. To meet the swiftly growing demands and to reduce the deployment time, datacenters use off the shelf commodity components such as processors. The foundation of these commodity components was laid out about half a century ago under completely different constraints and for different application characteristics. Back then, transistors were scarce, computation was much costlier than data movements, and memory size requirements were much less. Today, the computations are cheaper, the data movement has become the bottleneck, and applications have massive instruction footprints and datasets. Therefore, this project will lay a new architectural foundation for the future processors by specializing them for the modern and upcoming applications with new constraints and characteristics.

TailorMade will take a leap towards sustainable warehouse-scale computing by tailoring server processor microarchitecture to application characteristics for improving server request rate i.e., the number of requests served by a server per unit time. The inefficiencies in existing processor microarchitectures have pushed warehouse-scale computing towards the point of unsustainability exemplified by their energy consumption which is projected to account for about 8% of world's energy by 2030, compared to only 1% in 2010. Processors consume more than 60% of this energy; however, they sit idle for nearly 80% of time which not only results in significant energy wastage but also limits server request rate. The limited request rate overshoots the number of servers required to serve a clientele. The root cause of the limited request rate and high energy wastage in warehouse-scale computers (WSC) is the mismatch between the application characteristics and processor microarchitecture. Due to the massive (100s of MBs) instruction footprints of WSC applications, the front-end is unable to continuously supply instructions for execution, which limits performance and wastes energy. Also, due to the lack of instruction- and memory-level-parallelism (ILP and MLP), processors spend significant energy on executing instructions that do not contribute to performance and can potentially be executed cheaply. This project will tailor processor microarchitecture to WSC application characteristics for boosting server request rate while minimizing the energy wastage. To capture the massive instruction footprints, we will investigate co-design of instructions prefetching and replacement, as well as organization and code layout for instruction cache and branch target buffer. To address the lack of ILP and MLP, we will investigate segregating instructions based on their contribution to parallelism and investing energy-intensive resources exclusively in parallelism exhibiting instructions.

Funding scheme:

IKTPLUSS-IKT og digital innovasjon